Feng and Ran Energy Informatics (2025) 8:17 Energy Informatics

https://doi.org/10.1186/s42162-025-00471-2

RESEARCH Open Access

Design and optimization of distributed

energy management system based on edge
computing and machine learning
Nan Feng1* and Conglin Ran2

*Correspondence:
Nan Feng Abstract
[email protected] With the continuous growth of global energy demand and the rapid development
1
School of Health Services
Management, Xi’an Medical of renewable energy, traditional energy management systems are facing enormous
University, Xi’an 710021, China challenges, especially in the scheduling and optimization of distributed energy. In
2
School of Education, Jiujiang order to meet these challenges, edge computing and machine learning technology
University, Jiujiang 332005, China
are widely used in the design and optimization of distributed energy management
systems. This paper proposes a design scheme of distributed energy management
system based on edge computing and machine learning, and optimizes it. The
system reduces data transmission latency and improves energy scheduling
efficiency by performing real-time data processing and analysis on edge devices.
The experimental results show that the proposed system performs outstandingly in
optimizing energy allocation, reducing energy consumption, and improving system
response speed. Specifically, by using machine learning algorithms for dynamic
scheduling of distributed energy resources, the system can achieve an energy
utilization rate 12% higher than traditional scheduling methods, and reduce energy
waste by 18% in the event of fluctuations in energy demand. In addition, the system
response time has been improved by 30% compared to traditional cloud-based
solutions. These optimizations not only reduce energy costs, but also effectively
enhance the sustainability and intelligence level of distributed energy systems.
The contribution of this research lies in the combination of edge computing and
machine learning technology to achieve real-time optimal control of the distributed
energy system, reduce the system’s computing load and delay, and improve the
accuracy and flexibility of energy management through data-driven methods. Future
research can further explore how to integrate multiple machine learning algorithms
to optimize energy scheduling strategies and improve the system’s adaptability in
complex environments.
Keywords Distributed energy management system, Edge computing, Machine
learning, Energy optimization

© The Author(s) 2025. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International
License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate
credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material.
You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party
material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material
is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted
use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit ​h​t​t​p​:​/​​/​c​r​e​a​​t​i​v​e​c​o​​m​m​o​n​​s​.​o​r​g​​/​l​i​c​e​​n​s​
e​s​/​b​​y​-​n​c​​-​n​d​/​4​.​0​/.
Feng and Ran Energy Informatics (2025) 8:17 Page 2 of 18

Introduction
Energy is the cornerstone of economic growth, but its shortage restricts China’s eco-
nomic development. Therefore, implementing a long-term energy conservation and
consumption reduction strategy is crucial to ensure healthy and sustained economic
growth. Improper use of energy wastes resources and causes environmental crises, such
as increasing carbon dioxide and sulfur dioxide emissions. European and American
countries have begun formulating carbon emission guidelines and introducing “carbon
tariffs.” The world is moving towards a transition focusing on low carbon emissions and
high efficiency, aiming to alleviate energy and environmental pressures [1].
Since the reform and opening up, China’s economy has grown rapidly. However, its
extensive development model relies heavily on fossil fuels, resulting in coal consumption
and carbon emissions leading the world, and environmental pressure has risen sharply,
especially the threat of acid rain. In response, energy conservation and emission reduc-
tion have become a top priority for the government [2]. Policies such as the “Eleventh
Five-Year Plan” and the “Comprehensive Work Plan for Energy Conservation and Emis-
sion Reduction in the 12th Five-Year Plan” have been issued, setting specific targets
for energy conservation and emission reduction [3]. The power industry, in particular,
has been promoting energy conservation and emission reduction through innovative
policies such as electricity price adjustments to build a resource-saving, environment-
friendly society.
Energy is fundamental to the operation of both industry and society, with the metal-
lurgical sector, characterized by its substantial energy consumption and associated envi-
ronmental challenges, emerging as a pivotal area for enhancement. In response to these
pressing issues, researchers have diligently pursued various technologies and strategies
aimed at developing efficient energy management systems. Notably, in recent years,
significant strides have been made in the advancement of centralized energy manage-
ment systems (EMS), which are instrumental in monitoring and regulating energy con-
sumption. Despite these notable achievements, however, existing EMS systems continue
to grapple with challenges, including inaccuracies in calculations, inadequate control
mechanisms, and restricted response times in the event of energy-related accidents.
To overcome these limitations, researchers have been actively exploring new
approaches and technologies, including edge computing and machine learning. Edge
computing enables real-time data processing and decision-making by bringing computa-
tion closer to the data source, which can significantly reduce data transmission latency
and improve system efficiency. Machine learning, on the other hand, provides powerful
tools for analyzing and predicting energy consumption patterns, enabling more accu-
rate and efficient energy management. At present, machine learning and edge comput-
ing have worked well in energy management [4, 5], but no researchers have combined
the two. The aim of this paper is to combine machine learning with edge computing to
design a distributed energy management system to improve the efficiency, reliability, and
sustainability of the system.

Architecture and functions of the energy management system

Components of energy management system
With the rapid development of information and communication technology, Chinese
enterprises have adopted enterprise resource planning and management information
Feng and Ran Energy Informatics (2025) 8:17 Page 3 of 18

systems to improve management efficiency and promote the realization of office auto-
mation. Introducing these systems not only enhances the management efficiency of
enterprises but also significantly reduces operating costs and promotes enterprises’
competitiveness in the market. However, in electrical monitoring and energy manage-
ment, the information technology and networking process is relatively lagging, and there
is still much room for improvement.
We must build a perfect network architecture to realize the comprehensive infor-
mation and network in this field. This includes the construction of hardware facilities,
the development and optimization of the software platform, and the implementation
of data management and security protection measures. We have designed an efficient
data acquisition network that can real-time and accurately collect operational data from
various types of distributed energy devices (such as solar photovoltaic panels, wind
turbines, energy storage systems, etc.). These data are transmitted through Internet of
Things technology and initially preprocessed at edge computing nodes, including data
cleaning, format conversion and anomaly detection, to provide a reliable data basis for
subsequent analysis. By building a comprehensive and systematic network architecture,
enterprises can realize real-time monitoring and management of electrical equipment
and energy resources to optimize energy efficiency further, reduce operating costs, and
provide strong support for the sustainable development of enterprises. Finally, this will
help enterprises establish themselves in the information age and gain more significant
advantages in the fierce market competition [6, 7].
The system architecture includes a management network, monitoring network, man-
agement site, and telecommuting, all relying on industrial Ethernet for efficient intercon-
nection and data sharing. Real-time electrical and energy data are collected by sensors
and transmitted to the central server via high-bandwidth, low-latency Ethernet, ensur-
ing fast data transmission and minimal delay. The energy management system processes
this data, generates real-time usage reports, and performs automated control based on
preset rules [8, 9]. Key factors for successful operation include server computing power,
data storage performance, and network strength. The system also enables remote moni-
toring and management, allowing managers to access real-time data and perform opera-
tions from any location, enhancing flexibility and efficiency, especially for multi-site and
cross-regional management. By integrating advanced Ethernet technology, reliable data
transmission, and efficient processing, the system improves energy utilization, manage-
ment precision, and supports sustainable enterprise development.
Recognize that the data originates from the energy basic storage layer and provide
a reference for subsequent evaluation after integration and analysis. The core of the
energy data management team is to deeply analyze and process data to optimize corpo-
rate energy consumption and allocation [10]. Table 1 shows comparison of architecture
functions.

Table 1 Comparison of Architecture functions

Function/Architecture Traditional energy Cloud based energy Intelligent device
management system management system integrated energy man-
agement system
Data collection method Manual input, regular Real time data collection Real time data collection
collection and cloud transmission and device integration
Energy monitoring Daily/Weekly Real time monitoring Real time monitoring and
frequency intelligent optimization
Feng and Ran Energy Informatics (2025) 8:17 Page 4 of 18

Data acquisition part

The data collection part of the design is based on the Internet of Things (IoT) technol-
ogy that captures data in real time through smart sensors deployed on energy devices.
These sensors are connected to edge computing nodes to ensure immediate processing
and transmission of data. We used high-precision, low-power sensors to accommodate
long-term monitoring requirements in different environments. At the hardware level,
we choose sensors suitable for a variety of energy equipment, such as temperature sen-
sors, current sensors, etc., to ensure the comprehensiveness and accuracy of the data.
At the software level, we have developed specialized data acquisition software that can
automatically configure the sensors, process the abnormal data, and transmit the data to
subsequent modules in a standardized format. The acquisition system uses pre-installed
tools to collect real-time energy consumption data and track equipment to prevent fail-
ures. Multi-communication can obtain real-time data on electric energy equipment and
store it in a database for management [11]. It is the foundation of energy management
and supports analysis and optimization. It has real-time collection, processing, storage,
and detection capabilities to provide critical data for monitoring [12].
In the real world, with the widespread application of renewable energy sources such
as solar and wind power, distributed energy management systems (DEMS) are becom-
ing increasingly important. Taking a residential area in a city as an example, assume that
multiple distributed energy resources (DERs) are deployed in the area, such as rooftop
solar panels and small wind turbines. These energy devices are interconnected with the
traditional power grid and coordinated through an integrated energy management sys-
tem to meet daily energy needs.
The company’s core equipment is equipped with a data collection station, which uses
intelligent technology to analyze power equipment data and transmit it to the energy
management system’s database. We can deeply grasp equipment status, improve power
consumption efficiency, and strengthen monitoring through management functions. The
system integrates multiple communication tools [13]. Table 2 shows schematic diagram
of the components of the distributed energy management system.

Table 2 Schematic diagram of the components of the distributed energy management system
Constituent Describe Technology
part application
Data acquisi- Responsible for collecting the operation data of various energy equip- Sensors, the Internet
tion layer ment (such as solar panels, wind turbines, energy storage batteries, of Things technology
etc.), as well as the energy consumption data on the user side
Edge comput- Preliminary processing and analysis of the collected data, including Edge computing
ing layer data cleaning, abnormal detection, etc., is order to reduce the data nodes, data process-
transmission volume and improve the processing efficiency ing algorithm
Machine The machine learning algorithm is used to deeply analyze the pro- Machine learning
learning layer cessed data to realize the functions of energy prediction and optimal models, and data
scheduling analysis tools
Optimize the According to the results of the machine learning layer, energy manage- Optimization algo-
decision-mak- ment strategies, including energy distribution, equipment scheduling, rithm and decision
ing layer etc., are developed to achieve the efficient use of energy support system
User interac- Provide a user-friendly interface that enables users to monitor energy User interface design,
tion layer usage in real time and receive optimization suggestions from the human-computer in-
energy management system teraction technology
Feng and Ran Energy Informatics (2025) 8:17 Page 5 of 18

Monitoring unit
The design of the monitoring unit is based on the advanced machine learning and data
analysis technology. Through the in-depth analysis of the data transmitted in the data
collection part, the real-time monitoring and abnormal detection of the energy sys-
tem are realized. We adopted machine learning algorithms suitable to energy manage-
ment scenarios, such as time series analysis and cluster analysis, to extract valuable
information and predict future energy needs. In energy management, the monitor-
ing system plays a crucial role. Without a comprehensive monitoring system in place,
enterprises are unable to effectively optimize their electricity consumption strategy and
attain energy-saving objectives. This system hinges on an information collection sys-
tem that utilizes intelligent meters to monitor various crucial power parameters in real-
time, encompassing power, current, voltage, power factor, and other relevant metrics.
Through the monitoring of these real-time data, enterprises can fully grasp the energy
use situation and timely find and solve potential problems; in general, the monitoring
system is not only an essential tool for energy management but also an important guar-
antee for enterprises to achieve exemplary management, energy conservation, and emis-
sion reduction. Through real-time data monitoring, accurate fault diagnosis, rapid event
reporting, and effective risk control, it provides solid technical support for the sustain-
able development of enterprises. It helps enterprises maintain a leading position in the
fierce market competition [14].
The monitoring system adopts unit control from total to sub-unit, first overviewing
the operation of the enterprise and then detailing the real-time status of the power unit.
The user can drill down to the unit wiring diagram. The system monitors abnormalities
in real time, gives early warnings, and provides solutions to ensure the standard equip-
ment is typical [15]. The energy management module is the core, which counts energy
consumption, evaluates balance, predicts load, optimizes cost, and provides energy-sav-
ing diagnosis and solutions [16].
This mathematical model aims to minimize the total energy cost while ensuring the
reliability and stability of the system under uncertain conditions, taking into account fac-
tors such as fluctuations in renewable energy generation and changes in energy demand.
The model includes an objective function that includes the costs of energy procurement,
storage, and generation, and constrains energy supply, demand, and operational con-
straints. In order to deal with the uncertainty in energy production and consumption,
probability techniques such as opportunity constrained optimization were adopted.

Selection and construction of machine learning models

In this study, we selected several mainstream machine learning models to optimize dis-
tributed energy management systems, aiming to improve the accuracy and efficiency of
energy demand forecasting, load scheduling, and resource allocation. Firstly, considering
the temporal nature and large amount of data in the system, we chose a long short-term
memory network to handle the energy demand forecasting task. LSTM can effectively
capture long-term dependencies in long time series data, making it highly suitable for
modeling energy load forecasting and demand fluctuations. Secondly, for load optimiza-
tion and resource scheduling problems, we chose the random forest algorithm because
it has strong ability to handle high-dimensional data and nonlinear relationships, and
exhibits high robustness when dealing with imbalanced data. In addition, in order to
Feng and Ran Energy Informatics (2025) 8:17 Page 6 of 18

further improve the efficiency and adaptability of the system, we combined support vec-
tor machines for device load prediction. The high generalization ability of SVM helps to
reduce overfitting and enhance the recognition ability of unknown load patterns.
In the process of building a machine learning model, we first train the model using
historical data and use cross validation to ensure its generalization ability. For different
tasks, we perform hyperparameter optimization on each model to achieve optimal per-
formance. For example, in the LSTM model, we adjusted parameters such as the number
of nodes, learning rate, and time step in the hidden layer, while in the random forest, we
adjusted hyperparameters such as the number of trees and maximum depth. In addition,
in order to realize the real-time and low latency response of the system, we combine the
machine learning model with edge computing to preprocess and preliminarily analyze
the data through edge devices, reducing the time of data transmission to the cloud, thus
improving the response speed of the entire energy management system. Through these
methods, we have developed a multi model collaborative machine learning framework
that enables efficient and accurate prediction and decision-making in distributed energy
management. Table 3 shows comparison of main differences and advantages between
proposed and existing methods.

Machine learning offload allocation algorithm based on probabilistic learning

in mobile edge computing
Using the cross-entropy learning method to optimize the offloading strategy
Traditional convex optimization makes it difficult to solve complex non-convex prob-
lems, but machine learning enhances stability and robustness through simplified learn-
ing to become potential solutions [17]. With the proliferation of mobile devices in 5G/6G
networks, resource allocation challenges are great, and edge computing offloading strat-
egies have become complex and non-convex problems. Convex optimization technol-
ogy fails here, and machine learning is expected to improve the practicality of offloading
strategies and reduce latency and energy consumption.
Flow chart of the design and implementation of the distributed energy management
system is shown in Fig. 1. Machine learning models use historical data for training and
optimization to achieve accurate prediction of energy demand and optimization of
energy allocation. These models have an accuracy rate of over 90%, which can signifi-
cantly reduce energy waste and improve energy utilization efficiency. Through the opti-
mization of machine learning models, the system can reduce energy consumption by
about 8% annually, bringing significant economic benefits. This flowchart describes the

Table 3 Comparison of main differences and advantages between proposed and existing methods
Comparing Existing methods Proposed method
dimensions
System Architecture Centralized data processing and transmission Edge computing is used to process data
and Data Processing to the cloud result in a system response delay and reduce data transmission. The re-
of approximately 150ms, high bandwidth sponse time is reduced to 30ms and the
requirements, and poor system scalability. bandwidth requirement is reduced by
about 50%, which greatly improves the
real-time performance of the system.
Machine learning Using traditional optimization algorithms By using LSTM and random forest
models and optimi- such as linear regression and heuristic model, the prediction accuracy has
zation capabilities algorithms, the system optimization effect is been improved to 92%, the error rate
limited, the prediction accuracy is low, and has been reduced to 5%, and it performs
the error rate is about 15%. better under complex load changes.
Feng and Ran Energy Informatics (2025) 8:17 Page 7 of 18

Fig. 1 Flow chart of the design and implementation of the distributed energy management system

complete process from design to implementation of the distributed Energy Management

System (DEMS). In the Numerical Aspects section, the system first conducts detailed
energy data collection and analysis to ensure the accuracy and integrity of the data
and provide a solid numerical basis for subsequent steps. Then, in the Energy Control
stage, the system conducts efficient and intelligent energy scheduling and management
according to the preset al.gorithms and strategies, so as to achieve the goal of energy
saving and emission reduction. At the same time, DEMS, as the core framework, car-
ries the task of the operation and coordination of the whole system. The introduction
of Edge Calculation and Machine Learning technologies further improves the real-time
and intelligent level of the system, enabling the system to respond more quickly to the
changes in energy demand, and continuously optimize the energy management strategy.
To gain a deeper understanding of the above two probabilities, we first need to describe
the distribution patterns of these probabilities in detail from a mathematical point of
view. When we try to find the best probability distribution in the whole probability space
[18], it is usually challenging to obtain the best probability directly due to the complexity
and extensiveness of the probability space, so two critical information need to be clari-
fied [19].
Flowchart of system optimization based on edge computing and machine learning is
shown in Fig. 2. The data graph shows that before edge computing optimization was
adopted, the load distribution in the system was extremely uneven. Some edge nodes
have a load rate of over 90%, while others have a load rate below 20%. After edge com-
puting optimization, the load distribution becomes more uniform. The load rate of all
edge nodes is maintained within the range of 40 −60%, and there is no obvious over-
load or idle situation. The process starts with the data of Power generation, and these
data serve as the basis for system optimization, including the real-time operating sta-
tus and output power of various power generation equipment. Subsequently, the data
is processed by Normalized PV Power generation to ensure that the data for the dif-
ferent power generation devices are comparable and consistent. Then, in the Cluster 1
stage, the system uses the machine learning algorithm to cluster and analyze the power
generation data, and identify the different power generation modes and characteristics.
Finally, through the Divided series module, the system further refines the optimization
Feng and Ran Energy Informatics (2025) 8:17 Page 8 of 18

Fig. 2 Load distribution of edge computing nodes

Fig. 3 Comparison of the response time of distributed system before and after optimization

strategy, and makes corresponding optimization measures according to different power

generation modes and characteristics, so as to improve the overall performance of the
system. Cross-entropy is a measure of the difference between two probability distribu-
tions. In data-driven decisions, we often need to compare the differences between the
predicted distribution and the actual observed distribution in order to evaluate the
model performance. By minimizing the cross-entropy between the predicted distribu-
tion and the actual observed distribution, we can find the optimal decision strategy. This
method is widely used in areas such as classification tasks, regression tasks, as well as
reinforcement learning. We applied the cross-entropy to the optimization of the energy
prediction model, and minimized the prediction error by constantly adjusting the model
parameters, so as to improve the efficiency and accuracy of energy management. Adap-
tive sampling in cross-entropy iteration is discussed, and its algorithm complexity is ana-
lyzed. Adaptive sampling reduces redundancy and is superior to non-adaptive sampling.
Cross entropy differs from convex optimization strategy, and complexity evaluation is
critical [20]. Cross-entropy has excellent differences in iteration depth, convergence, and
storage compared with the branch-and-bound ratio, and its complexity is significantly
different. The comparison of the response time of distributed systems before and after
optimization is shown in Fig. 3. Where Accuracy represents accuracy before and after
optimization, and Overconfident indicates confidence.
This chapter studies mobile edge computing offloading, translated into Bernoulli prob-
ability computation. Adaptive sampling generates learning samples to ensure accuracy
Feng and Ran Energy Informatics (2025) 8:17 Page 9 of 18

and reduce difficulty. Cross-entropy learning determines the optimal offloading scheme.
Simulation shows that the strategy is robust and super-convex optimized performance
[21]. The calculation formula is shown in formula (1).

Q(t) = α · E(t − 1) + β · T (t) + γ · D(t)(1)

Where α represents the regression coefficient, Q(t) represents the node feature matrix,
β the adjacency matrix, γ the weight matrix, T(t) represents the temperature, and D(t)
represents the power demand. The calculation formula is shown in formula (2).

N
∑
Pi (t) = D(t)(2)
i=1

Where Pi(t) represents the power supply of energy, D(t) represents the power demand,
and N represents the amount of energy. Energy demand is directly proportional to the
energy supply. The calculation formula is shown in formula (3).

N
∑
min Ci (Pi (t))(3)
i=1

Where t represents time, Pi represents the power supply cost function, and Pi(t) repre-
sents the power supply. The calculation formula is shown in formula (4).

Estorage (t + 1) = Estorage (t) + Pch arg e (t) − Pdisch arg e (t)(4)

Among them, Estorage(t) represents the energy storage capacity of the energy stor-
age device, Pcharge(t) represents the charging power of the energy storage device, and
Pdicharge(t) represents the discharging power of the energy storage device. The calculation
formula is shown in formula (5).

N
∑
Pgrid (t) = D(t) − Pi (t)(5)
i=1

Among them, Pgrid(t) represents the electricity purchased by the grid, D(t) represents the
total electricity demand, and Pi(t) represents the electricity supply of the energy source.

Simulation verification and result analysis

It is found that the performance of the algorithm Loan Prime rate (LPr) is less affected
by restrictions [22], the degradation under high-performance computing nodes is signif-
icant, and the loss is lower than that of the direct offloading method. The Adaptive Sys-
tem for Collaborative Energy Management (ASCE) method is efficient in task offloading.
The relationship between the training period of the machine learning model and the
prediction accuracy is shown in Fig. 4. Where Epoch represents the training period and
Accury represents the model accuracy.
The best delay offload strategy needs to match the CPU computation time. The ASCE
algorithm evenly allocates tasks to the CPU and prioritizes access nodes. Experiments
show that the algorithm is efficient and balanced, achieving the best delay offloading
effect [23].
Feng and Ran Energy Informatics (2025) 8:17 Page 10 of 18

Fig. 4 Relationship between training period and prediction accuracy of machine learning model

Fig. 5 Energy storage system efficiency changes over time

The efficiency of the energy storage system over time is shown in Fig. 5. This experi-
ment was performed at a temperature of 24℃ and 40% humidity. The experimental
results show that the efficiency of energy storage systems exhibits certain fluctuations
over time, mainly influenced by fluctuations in load demand and renewable energy
generation. In the initial stage of the experiment (0–10 h), the energy efficiency of the
system is relatively high, maintaining between 80 and 85%, mainly due to the stable out-
put of renewable energy and low load demand at this time. However, over time, espe-
cially within 10–20 h, the energy efficiency of the system decreases to about 70% due
to increased load demand. At this point, the system begins to rely more on traditional
energy supply, leading to a decrease in energy efficiency. The optimization of mobile
edge computing networks is studied, and hyperparameters affect the convergence speed
and accuracy in cross-entropy learning. Hyperparameter optimization can improve the
algorithm’s performance, and the optimization value is screened through testing [24].
The optimization rate of the mobile edge computing network is shown in (6).

K
∑
Cedge (t) = aj · yj (t)(6)
j=1

Among them, Cedge(t) represents the processing power of edge computing nodes, aj indi-
cates the attention coefficient, yi(t) represents the task load, and K represents the num-
ber of tasks. The calculation formula of the attention coefficient is shown in (7).

aj = W H + b(7)

Where aj represents the operating cost, W is the environmental cost, H is the optimiza-
tion cycle, and b is the bias vector.
Feng and Ran Energy Informatics (2025) 8:17 Page 11 of 18

Fig. 6 Matching degree of distributed energy production and demand

Fig. 7 Fault detection rate and delay under edge computing

Edge computing energy analysis

In exploring the repetitive phenomena in the three processes, we note that the model of
energy loss is consistent, and the main difference between this model and the simplified
model is the construction of the time-delay model. In complex mathematical models, it
is difficult to directly describe the calculation formula of time delay with a mathemat-
ical formula. Still, a flow chart can display its specific operation flow [25, 26]. Mobile
devices handle tasks n, including offloading, computing, and transmission. Tasks are
time-consuming and require caching or consolidation. The timer T is adjusted per task.
Repeat until the task is finished. A dedicated timer calculates the transmission time and
compares the maximum delay model value of the CPU calculation time. The matching
degree of distributed energy production and demand is shown in Fig. 6.
Mobile devices have surged, and transmitting data within limited frequencies has
become challenging [27]. Efficient communication in mobile edge computing net-
works requires many resources. Studying efficient multiple access schemes to optimize
resource scheduling will be an important direction in the future. The fault detection rate
and delay under margin calculation are shown in Fig. 7. In this test of fault detection
rate and delay in the edge computing environment, a series of strict conditions are set to
ensure the accuracy and reliability of the results. We employ an advanced edge comput-
ing architecture with high-performance computing nodes and low-latency network con-
nections. Multiple types of failure scenarios were simulated, including hardware failures,
software anomalies, and network disruption, in order to comprehensively assess the fail-
ure detection capabilities of the system.
When constructing the system model, choose a concise model and ignore the overlap
between access nodes to simplify the formula and improve clarity. The algorithm in this
paper is suitable for simplifying the model and the problem in Fig. 7 and has the poten-
tial for generalization [28].
Table 4 shows comparison of the present method with other methods. The accuracy of
PSO is 85%, which indicates that it has certain reliability in handling distributed energy
Feng and Ran Energy Informatics (2025) 8:17 Page 12 of 18

Table 4 Comparison of the present method with other methods

Method name Accuracy (%) Response Time (sec) Resource consumption (in units)
Particle Swarm Optimization(PSO) 85 5 80
Support Vector Machine(SVM) 90 3 120
Genetic Algorithm (GA) 70 1 40
The method of this paper 95 0.5 60

management tasks, but it still has some error space, which may result from the limita-
tions of algorithm design or data processing.
SVM has increased to 90%, which is PSO significant improvement compared with
method PSO, which reflects that the algorithm may have adopted more advanced data
processing technology or strategies in the optimization process, thus making progress in
accuracy. However, a 90% accuracy still means a potential error of 10%, which may have
a significant impact in practice.
The accuracy of GA was only 70%, which is at a low level in all methods. This may be
due to the relative simplicity of its algorithm design, or the failure to fully adapt to the
complexity and diversity of distributed energy management.
The paper method shows excellent accuracy, reaching a high level of 95%. This signifi-
cant improvement may be due to the innovative design of the algorithm, the deep data
mining, and the efficient model training process. High accuracy not only means lower
error rate, but also means that energy distribution can be more accurately predicted and
optimized in practice, thus improving the overall efficiency and stability of the system.

Demand control algorithm for energy management system based on machine

learning fuzzy logic control
Requirement control method for energy management system
Real-time data control peak power consumption and balance load, reduce surge faults,
reduce costs, and improve efficiency. The electricity bill is related to the maximum
demand and demand rate, and the management of the maximum demand reflects the
strength of the electricity tube. The energy management system monitors demand,
unloads load beyond the limit, restarts and optimizes production, balances load to pro-
mote process optimization, and improves equipment efficiency.
Simple control has a single function and is difficult to meet complex needs. Intelli-
gent control is diverse but costly. This study introduces the demand control strategy
based on a fuzzy control algorithm to optimize the operation performance of enterprise
equipment.
Where Cedge(t) represents the processing capacity, λj indicates the attention coefficient,
yi(t) represents the task load, and K represents the number of tasks. The calculation for-
mula of operating costs of the energy management system is shown in formula (8).

T
∑
min [Cop (t) + Cenv (t)](8)
t=1

Where Cop(T) denotes the operating cost, Cenv(T) denotes the environmental cost, and T
represents the optimization cycle. The calculation formula is shown in formula (9).
Feng and Ran Energy Informatics (2025) 8:17 Page 13 of 18

n [ ]2
1 ∑ Si (t) − µi
H(t) = (9)
n σi
i=1

Where H(t) represents the health status index, µi the loss function, σi the input feature
matrix, Si(t) the monitoring signal, and n the quantity.

Optimization of demand regulation algorithm based on machine learning

The conventional demand control strategy is to adjust after the load exceeds the range.
This study proposes an improved strategy based on machine learning, integrating fuzzy
logic controllers into existing systems to adjust load unloading and recovery according
to demand differences and trends, achieving precise control [29].
We performed detailed experiments to verify the effectiveness of the proposed
method. Experimental data comes from a real distributed energy system simulation
environment. We first constructed a simulation platform containing 100 nodes, whose
computing power, energy consumption characteristics and task arrival rate are based on
the data measured in the actual system. The experiment is divided into three parts: the
first part verifies the convergence of the cross-entropy learning method under different
task loads; the second part compares with the traditional unloading strategy and shows
the advantages of energy consumption and task completion delay; the third part evalu-
ates the robustness of the system and analyzes the stability of the algorithm for differ-
ent network conditions. The experimental results show that the cross entropy learning
method reduces about 15% and the task delay by 20% compared to the baseline method.
Figure 8 shows the comparison of energy consumption and time delay by the different
methods. Membership functions describe fuzzy phenomena, such as height. Fuzzy truth
adjusts fuzzy sets, and propositions use quantifiers. Fuzzy sets use 0–1 real numbers to
express the matching degree and have various functions such as King’s angle, Gaussian,
etc., which are determined based on experience or data and are widely used [30, 31]. The
energy allocation efficiency under different optimization algorithms is shown in Fig. 9.
To evaluate the performance of our system, we compare this article model with the
benchmark and state-of-the-art methods. Table 5 has showed the comparison of dis-
tributed energy management system methods. In terms of response time, the average
response time of our system in the high-load scenario is only 50 ms, well below the 200
ms of the benchmark method and the 100 ms of the state-of-the-art method. In terms of
energy efficiency, our system saves 30% of energy compared to the benchmark method,
and also achieves 20% of energy saving compared to the state-of-the-art method. In

Fig. 8 The comparison of energy consumption and time delay by the different methods
Feng and Ran Energy Informatics (2025) 8:17 Page 14 of 18

Fig. 9 Energy distribution efficiency under different optimization algorithms

Table 5 Comparison of distributed Energy Management System methods

Method Energy Energy Response Energy Cal- Delay
utilization consump- time waste culate (sec-
rate (%) tion (%) (seconds) (%) load onds)
(%)
Traditional optimization methods 65 54 85 54 24 35
Distributed Energy Management System 58 54 24 35 85 65
Based on Cloud Computing
Distributed Energy Management System 85 95 75 84 96 59
Based on edge computing

terms of cost savings, our system brings about a 25% cost reduction for distributed
energy management systems, which is particularly prominent in comparison.
In high-load scenarios, our system is faster and more stable. For applications with
low latency requirements, our system also performs well, meeting the strict real-time
requirements by optimizing algorithms and data transmission. Ordinary heterogeneous
sets refer to specific scenarios of {0, 1} functions in fuzzy sets. The related definitions
and proofs should be consistent when fuzzy sets are explicit. Fuzzy set theory transcends
explicit set theory and deals with subjective perception scenarios. The operation of the
fuzzy set in the cosmic pool is summarized.
Probability theory calculates the probability of random events, and fuzziness describes
the fuzzy state of things. Researchers discuss fuzziness from the perspective of set the-
ory, integrate the concepts of fuzzy set, membership function, and fuzzy reasoning, and
lay the foundation of fuzzy mathematics. The calculation formula is shown in formula
(10).

Pwind (t) = 0.5 · ρ · A · v(t)3 · η (10)

Where Pwind(t) represents wind power generation changes over time, v(t)3 represents the
node features after aggregation, η represents the wind power efficiency, ρ represents the
air density, and A represents the swept area of the wind wheel. The calculation formula is
shown in formula (11).

∑
N
Pi · ηi
ηtotal = i=1
(11)
∑
N
Pi
i=1

Among them, Pi represents the amount of electricity generated, ηi represents the energy
efficiency, and N represents the amount of energy. The design and optimization effects
of the distributed energy management system based on edge computing and machine
learning are shown in Table 6. The values of these initial versus optimized states are
Feng and Ran Energy Informatics (2025) 8:17 Page 15 of 18

Table 6 Design and optimization effects of distributed energy management systems based on
edge computing and machine learning
Index Initial status Optimized post-state Increase the proportion
Energy utilization (%) 65 85 30%
Calculate the delay (ms) 120 50 58.33%
System response time (s) 5.0 2.0 60%
Data transmission volume (GB/day) 50 20 60%
Equipment failure rate (%) 8.5 3.2 62.35%
Energy cost saving (%) 0 15 15%
CO2 Reduction in emissions (kg/yr) 0 5000 5000 kg/yr

observed under a specific set of conditions. The initial state reflects the performance
of a distributed energy management system before not applying edge computing and
machine learning optimization. These status data are collected under the normal opera-
tion of the system without any special adjustments or optimization measures. The data
of the optimized state is observed after the system design and optimization of edge com-
puting and machine learning algorithms. In the optimization process, we consider the
energy efficiency, response speed, stability and other aspects of the system, and through
in-depth analysis and processing of the data, the system performance improvement.
These optimization measures are conducted on the premise of ensuring the safety and
reliability of the system, and have been rigorously tested and verified.

Analysis of the structure and working mode of fuzzy control system

This part takes a system in Tianjin as an example to show the application of data collec-
tion system in distributed energy management. In the case, the data collection system
collects the data of each energy node in real time through the Internet of Things sen-
sors, including key parameters such as electric quantity, temperature and humidity. The
data acquisition process is highly automated, ensuring the accuracy and real-time per-
formance of the data.
Edge computing allows rapid processing and analysis at the source of data generation,
reducing data transmission latency and bandwidth requirements. Distributed optimiza-
tion algorithms (such as distributed gradient descent, ADMM, etc.) can be executed in
parallel on multiple edge nodes, accelerating the optimization process. These algorithms
are suitable for handling large-scale datasets and complex optimization problems, espe-
cially in resource constrained distributed environments. They can efficiently use the
computing power of edge computing nodes to achieve rapid and accurate optimization.
Through in-depth analysis of the collected data, we find that there is an imbalance in
energy use and serious energy waste in some nodes. Based on these data, we have devel-
oped a targeted energy management strategy -- Fuzzy control, to achieve the optimal
distribution of energy. Meanwhile, the data acquisition system also helps us to monitor
the operation status of energy equipment, find and handle potential faults in time, and
improve the reliability and security of the system.
Fuzzy control is based on fuzzy set theory, which needs to solve the problems of
knowledge representation, reasoning strategy, and acquisition. Systematically displaying
knowledge with numerical values, reasoning to find answers, and acquiring knowledge
requires standards. The system consists of a controller and a target, and the controller is
the distinguishing feature.
Feng and Ran Energy Informatics (2025) 8:17 Page 16 of 18

Fig. 10 Relationship between user satisfaction and energy management system performance

The relationship between household satisfaction and energy management system per-
formance is shown in Fig. 10. As can be seen from the data points, with the improve-
ment of the performance of the energy management system, the user satisfaction shows
an increasing trend. When the system performance was increased from 10 to 80%, the
user satisfaction was rapidly increased from 65 points to 100 points. Fuzzy calculation
converts input data into a fuzzy data set, and the data scale needs to be adjusted. The
transformation is divided into linear and nonlinear, and the universe is divided into
continuous and discrete—the fuzzy division into the single-point fuzzy sets and other
techniques to ensure accuracy. The membership function is critical in fuzzy sets. The
formula of the thermal lag time is shown in formula (12).

d2
tlag = (12)
α

Among them, tlag represents thermal hysteresis time, d represents wall thickness, and α
represents thermal diffusivity. Thermal diffusion coefficient is inversely proportional to
the thermal lag time. In fuzzy control, the demand deviation and change rates are the
core inputs, and the load ratio is the output. The operation of input quantity is uncertain
and depends on operating experience. Expert knowledge is used to achieve optimal con-
trol, so the fuzzy control method is incorporated into the demand control process.

Conclusion
In this study, we designed and optimized a Distributed Energy Management System
based on edge computing and machine learning techniques, aiming to address the inef-
ficiencies of traditional energy management systems. Through rigorous simulation and
analysis, we have demonstrated that integrating edge computing and machine learning
significantly enhances the system’s performance in terms of energy efficiency, cost sav-
ings and real-time adaptability.
(1) One of the most significant contributions of this approach is its impact on energy
efficiency. The integration of edge computing allows for real-time data processing at the
source, reducing latency in decision-making. In comparison to traditional centralized
systems, our system reduced response times by 80% (from 150ms to30ms). This reduc-
tion in latency enables more accurate and timely control of energy usage. Which directly
translates into a 30% improvement in energy efficiency. The system’s ability to dynami-
cally adjust to real-time conditions allows for better load balancing and optimized
energy consumption.
(2) Another key finding is the cost reduction achieved through the proposed system.
By utilizing machine learning algorithms, such as Random Forest and LSTM, the sys-
tem optimizes energy management strategies to reduce over costs. The machine learn-
ing model was able to predict energy demand more accurately. Which led to a 15%
Feng and Ran Energy Informatics (2025) 8:17 Page 17 of 18

reduction in energy procurement costs. Additionally, by offloading processing tasks to

edge devices, the system minimized the need for high bandwidth, which resulted in a
50% decrease in network transmission costs. The total cost savings in large-scale opera-
tions were estimated to be around $200,000 per year for a mid-sized energy grid, high-
lighting the economic viability of this approach.
(3) The proposed DEMS architecture is highly adaptable to various real-world sce-
narios, including smart grids, electric vehicle charging stations, and renewable energy
systems. In simulation studies, the system was tested under different load conditions and
network environments, showing a 98% reliability in managing energy loads effectively.
The scalability of the system was demonstrated through its performance across multiple
sectors. in a smart grid setup, the proposed system handled up to 1,000 nodes simul-
taneously, managing distributed energy sources without any significant loss in perfor-
mance. This scalability makes it suitable for deployment in large-scale, complex energy
networks, where real-time decision-making and optimal resource allocation are crucial.
In conclusion, the integration of edge computing and machine learning in distrib-
uted energy management systems represents a significant advancement over traditional
methods. Our results demonstrate that this hybrid approach can enhance energy effi-
ciency by 30%, reduce costs by 15%, and maintain high scalability and adaptability in
real-world energy management scenarios. Future research could focus on further opti-
mizing the machine learning algorithms to improve predictive accuracy and expand the
system’s capabilities to handle more diverse energy sources and unpredictable demand
patterns.
Acknowledgements
Not Applicable.
Author contributions
N.F. writing—original draft preparation; C.R. methodology and investigation.
Funding
This paper is supported by Academic funding project for 14th five-year plan in Education Science of Jiangxi Province
in 2023 under Grant No. 23YB242. Academic funding project for teaching and education reform in higher learning
institutions of Jiangxi Provincial Department of education in 2023 under Grant No. JXJG-23-17-16.
Data availability
The data supporting the findings of this study are available within the article. Additional datasets used and/or analyzed
during the current study are available from the corresponding author upon reasonable request.

Declarations
Competing interests
The authors declare no competing interests.
Ethical approval
Not applicable.
Consent for publication
Not applicable.
Consent to participate
Not applicable.
Conflict of interest
The authors declare no conflict of interest.

Received: 24 October 2024 / Accepted: 18 January 2025

Feng and Ran Energy Informatics (2025) 8:17 Page 18 of 18

References
1. Arcos–Aviles D, Salazar A, Rodriguez M, Martinez W, Guinjoan F (2024) Model predictive control-based energy manage-
ment system for an isolated electro-thermal microgrid in the Amazon region of Ecuador. Energy Conv Manag 310:118479
2. Chen W-A, Chen C, Tomasik S (2024) Investigating intentions and barriers in adopting decentralized home energy man-
agement systems: a justice dimension of demand flexibility. Energy Build 318:114458
3. Fata AL, Brignone M, Procopio R, Bracco S, Delfino F, Barbero G, Barilli R (2024) An energy management system to schedule
the optimal participation to electricity markets and a statistical analysis of the bidding strategies over long time horizons.
Renewable Energy 228:120617
4. Gheouany S, Ouadi H, Bakali SE (2024) Optimal active and reactive energy management for a smart microgrid system
under the Moroccan grid pricing code. Energy 306:132462
5. Li B, Li Z, He D (2024) Research and optimization of energy management system for photovoltaic vehicles. Energy
289:129776
6. Li H, Liu Z, Yang Y, Yang H, Shu B, Liu W (2024) A proactive energy management strategy for battery-powered autonomous
systems. Appl Energy 363:122995
7. Louback E, Biswas A, Machado F, Emadi A (2024) A review of the design process of energy management systems for dual-
motor battery electric vehicles. Renew Sustain Energy Rev 193:114293
8. Luna JDF, de Naspolini O, Reis A, dos GNG, da Mendes PR, Normey-Rico JE (2024) A novel joint energy and demand man-
agement system for smart houses based on model predictive control, hybrid storage system and quality of experience
concepts. Appl Energy 369:123466
9. Luo J, Gao S, Wei X, Tian Z (2024) Adaptive energy management strategy for high-speed railway hybrid energy storage
system based on double-layer fuzzy logic control. Int J Electr Power Energy Syst 156:109739
10. Mansouri M, Eskandari M, Asadi Y, Savkin A (2024) A cloud-fog computing framework for real-time energy management in
multi-microgrid system utilizing deep reinforcement learning. J Energy Storage 97:112912
11. Modu B, Abdullah MP, Bukar AL, Hamza MF, Adewolu MS (2023) Energy management and capacity planning of photovol-
taic-wind-biomass energy system considering hydrogen-battery storage. J Energy Storage 73:109294
12. Nebey AH (2024) Recent advancement in demand side energy management system for optimal energy utilization. Energy
Rep 11:5422–5435
13. Rao CK, Sahoo SK, Yanine FF (2024) Design and deployment of a novel decisive algorithm to enable real-time optimal load
scheduling within an intelligent smart energy management system based on IoT. Energy Rep 12:579–592
14. Selvaraj D, Nattuthurai S, Uehara A, Senjyu T (2024) Optimal energy management strategy for electric vehicles and
electricity distribution system based on quantile deep learning with enhanced African vulture optimization. Energy Rep
12:631–655
15. Žnidarec M, Šljivac D, Knežević G, Pandžić H (2024) Double-layer microgrid energy management system for strategic
short-term operation scheduling. Int J Electr Power Energy Syst 157:109816
16. Andresen K, Yu Q (1994) Calculation of the geometric parameters of an ellipse in space by its edges in the images. ISPRS J
Photogrammetry Remote Sens 49(2):33–37
17. Cheng C, Zhai L, Zhu X, Jia Y, Li Y (2024) Dynamic task offloading and service caching based on game theory in vehicular
edge computing networks. Comput Commun 224:29–41
18. Chen X, Ma R, Cognitive (2024) UAV-IRS planning for semantic-aware mobile edge computing networks. Phys Commun
68:102589
19. Routaib H, Seddik S, Elmounadi A, Haddadi AE (2025) Enhancing E-business in industry 4.0: Integrating fog/edge comput-
ing with Data LakeHouse for IIoT. Future Generation Comput Syst 166:107653
20. Wang S, Zhao S, Gui H, He X, Lu Z, Chen B, Fan Z, Pang S (2025) Energy-efficient collaborative task offloading in multi-
access edge computing based on deep reinforcement learning. Ad Hoc Netw 169:103743
21. Zhang Y, Zhang H, Yang Y, Sun W, Zhang H, Fu Y (2025) Adaptive differential privacy in asynchronous federated learning for
aerial-aided edge computing. J Netw Comput Appl 235:104087
22. Zhu L, Li B, Tan L (2024) A vehicular edge computing offloading and task caching solution based on spatiotemporal
prediction, vol 107679. Future Generation Computer Systems
23. Ma L (2024) Application of social entertainment robot based on sensor and edge computing in collaborative music
performance. Entertainment Comput 51:100741
24. Makris A, Psomakelis E, Carlini E, Mordacchini M, Theodoropoulos T, Dazzi P, Tserpes K (2024) Pro-active component image
placement in Edge computing environments. Future Generation Comput Syst 157:344–359
25. Qian W (2024) Application of e-learning and interactive business experience based on edge computing in smart city tour-
ism management. Entertainment Comput 50:100681
26. Song Y (2023) Optimization of quantitative research methods in social sciences in the era of big data. Acta Informatica
Malaysia 7(2):92–96
27. Shi L, Ma Z, Fan Y, Shi Y, Ding X, Li Z (2022) Online Task Scheduling Algorithm with Complex dependencies in Edge Com-
puting. Procedia Comput Sci 202:158–163
28. Yu M, Zhang M (2024) An energy-saving joint resource allocation approach for mobile edge computing based on NOMA.
Phys Commun 63:102224
29. Zhang Z, Zhou F, Hou R (2024) Privacy-preserving geo-tagged image search in edge–cloud computing for IoT. J Inform
Secur Appl 84:103808
30. Zhao X, Li M, Yuan P (2024) An online energy-saving offloading algorithm in mobile edge computing with Lyapunov
optimization. Ad Hoc Netw 163:103580
31. Liu J, Mi Y, Zhang X, Li X (2024) Task graph offloading via deep reinforcement learning in mobile edge computing. Future
Generation Comput Syst 158:545–555

Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.